In have two previous posts, I sought to combine coffee and deep learning: this one and the latest. Anyone who has been following my blog for a while knows I start with a beginners mind. Lol. In other words, I get almost everything wrong for a while until I sort it out. I appreciate that folx are willing to read along and humor me. My friend and colleague Piero Toffanin (founder of WebODM and core developer of OpenDroneMap) might have called me out gently in a message, but so gently I missed it.
Mistakes were made…
So here was my supposed success:
:Lolsob: Ok. Now that I know how to read these a little better, this was a terrible model. Back to the drawing board. I did a few things. As per the previous post, I did some 3D renderings of the data, with rotations and translations in space. All of the following are the same 2.3 bars of pressure:
Honestly, I did this before I knew my previous model was so poorly trained, just as an excuse to train deep learning with 3D data from blender. But, it worked out great! Keep in mind, though, I am still a noob, so mistakes may still be extant.
And now we work on refining by retraining some of our deeper layers (I don’t know how many gauges show up in resnet50:
Training loss is down to 0.59 bars and validation loss down to 0.04 bars!
Wow! Ok. Time to wire this up to a web app and get a proof of concept working. Can I do this just with and HTML5 app with camera access? Time will tell. I hope so: I’m not sure I have bandwidth for much more than that… .
One thought on “Bean 3 — Gauging the best way to combine Deep Learning and Coffee”
Did you ever build a model or web app for this? Been toying around with my own ML for coffee and this seems like it could be more successful than what I had tried.